Question about the Federated setting in Track A

I saw that the dataset is vertically partitioned and, correct me if I’m wrong, that only SWIFT has access to the labels.

  1. This means that all parties (SWIFT, and different Banks) have different types of data and each cannot train on a similar architecture like we would in regular/Horizontal Federated Learning, is that correct ?
  2. If so, as Banks have no access to the transaction information held by SWIFT, is there such a thing as a ‘local model’ for banks ?
  3. How are you gonna test the accuracy at test/inference time ? Is test time ‘private’, ie. are we gonna use the exact flags from the banks ?

Thank you!

Hi @philippe.liu,

  1. Correct—teams will need to account for the vertical partitioning as part of their solution architectures.
  2. Each federation unit (SWIFT data holder + bank data holders) will be capable of running computation, and the aggregator is capable of running computation during training. The particulars of what the bank nodes run is up to your design for your solution.
  3. For testing, there will be a test split of transaction data available to the SWIFT node. The bank nodes will have the same account data.